Tapping into new ‘probabilistic computing’ paradigm can make AI chips use much less power, scientists say
American and Japanese scientists have used a new type of component in artificial intelligence (AI) chips that consumes less power when performing advanced calculations. The new system allows more operations to run in parallel, allowing the chip to achieve peak performance more efficiently.
The majority of computers rely on bits – the 0s and 1s that represent digital information and that programs use to carry out instructions – but some specialized technologies, such as neuromorphic chipsuse probabilistic bits (p-bits) instead.
Although the randomness of p-bits is useful, developers still need to control how often they produce a 0 or a 1 so they can guide their system to better answers. Most p-bits are therefore built with digital-to-analog converters (DACs), which use analog voltages to bias them in one direction or the other. But these are bulky and consume a lot of energy.
“The reliance on analog signals was holding back progress,” said the study co-author. Shunsuke Fukamiprofessor in materials science, in a statement. “So we discovered a digital method to adjust the behavior of the p-bits without the need for the large, bulky analog circuits typically used.”
Instead of DACs, the scientists constructed their p-bits using magnetic tunnel junctions (MTJs) – tiny devices that naturally switch between 0 and 1 randomly – and fed this stream of bits into a local digital circuit. Depending on how long the circuit waits to combine these random 0s and 1s, and how it counts and weighs each, the final output p-bits may become mostly 0s or 1s.
The scientists presented their results in a study published on December 10, 2025 in 71st International Meeting on Electronic Devices in San Francisco. The work was carried out in collaboration with Taiwan Semiconductor Manufacturing Company (TSMC), the world’s largest semiconductor foundry.
Circuit parameters can be adjusted by a user or a program, allowing control to what extent the p bit favors a value. Above all, because this control is entirely digital, it requires much less space and power on the chip than traditional DACs.
Self-organized behavior adds to efficiency
Another advantage of the new approach is that p-bits can demonstrate “self-organized” behavior, the scientists said. With DACs, when a user specifies a preference for mostly 1s or 0s, an analog signal continually biases the p-bits. They all feel this push at the same time, which creates the risk that they will all produce an outcome simultaneously.
Ideally, the p-bit outputs would be produced in a staggered manner, so that they have the opportunity to read the outputs of the previous p-bits and use that information to decide whether going to 0 or 1 will be more useful for the overall calculation.
With the new system, when the user adjusts the parameters for the desired bias, a digital signal is sent to the local control circuit of each p-bit. Since each circuit generates its subsequent output using its own unique timing, the p bits naturally avoid updating at the same time. Staggered outputs also allow multiple p-bits to operate in parallel and explore multiple possible solutions at once, allowing chips to perform calculations more efficiently.
Until now, the costs of using DACs have prevented the mass production and use of p-bits in commercial AI hardware, but this breakthrough could change that, scientists say. Efficiency gains can help reduce significant environmental impact of current AI systems.
The team behind the MTJ-based p-bits has yet to release performance benchmarks compared to conventional DAC designs, meaning it is uncertain how feasible commercialization is at this point. Thermal stability and reliability while controlling switching current are known challenges for MTJs. Nonetheless, the team is optimistic that their energy breakthrough will make probabilistic computing more accessible in other areas, including solving routing problems in logistics and rapidly exploring large numbers of solutions in scientific discovery.
